Comparison SARIMA-LSTM and CNN-LSTM Machine Learning Algorithms in Predicting Time Series Analysis: Road Traffic Accidents Data

Abstract

Traditional time series models may not adequately capture the underlying patterns for prediction in particular time series data. Hybrid machine learning approaches offer a more effective solution. This study compares two hybrid machine learning approaches, Seasonal Autoregressive Integrated Moving Average (SARIMA)-Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN)-LSTM—for time series forecasting of historical road traffic accident counts. The SARIMA-LSTM model combines the ability of SARIMA to capture trend and seasonality with the nonlinear learning capacity of LSTM networks, while the CNN and LSTM model integrates convolutional feature extraction with temporal sequence learning. Time series road traffic accident counts spanning five consecutive years (1399–1403 Iranian calendar) were used to evaluate both models. The findings demonstrate that the CNN-LSTM model consistently outperforms the SARIMA-LSTM approach across all evaluation metrics, achieving significantly lower error rates and providing more reliable forecasts for traffic accident prediction.